2021/06/15 by Stefan Grünewald, Grünewald, Stefan
Computer Science · #Computation and Language (cs.CL) #FOS: Computer and information sciences #Natural Language Processing Techniques #Software Engineering Research #Text Readability and Simplification #Topic Modeling
paper · pdf · doi:10.48550/arxiv.2106.08159
openalex publication_date 2021/06/15 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
Modern graph-based syntactic dependency parsers operate by predicting, for\neach token within a sentence, a probability distribution over its possible\nsyntactic heads (i.e., all other tokens) and then extracting a maximum spanning\ntree from the resulting log-probabilities. Nowadays, virtually all such parsers\nutilize deep neural networks and may thus be susceptible to miscalibration (in\nparticular, overconfident predictions). In this paper, we prove that\ntemperature scaling, a popular technique for post-hoc calibration of neural\nnetworks, cannot change the output of the aforementioned procedure. We conclude\nthat other techniques are needed to tackle miscalibration in graph-based\ndependency parsers in a way that improves parsing accuracy.\n